Every AI session starts with amnesia. Here’s how to fix it.
If you’ve been using Claude for client work, you already know the tax: every new session, you spend the first five minutes re-explaining who the client is, what their brand sounds like, what you’re trying to accomplish, and what not to do. It’s repetitive, it’s slow, and it means the first output of every session is always the worst one.
If you’re new to Claude Code, start with the first post in this series — it covers what Claude Code is and how to get it running. This post assumes you have it installed.
There’s a file that changes this. It’s called CLAUDE.md, and it’s one of the most underused features in Claude Code.
Here’s how it works, and how to set one up for every client in your agency.
What Is CLAUDE.md?
CLAUDE.md is a plain Markdown file that Claude Code reads automatically every time you start a session in a given project folder. You don’t have to paste it, mention it, or ask Claude to read it. The moment you open Claude Code in a directory that contains a CLAUDE.md file, Claude ingests it as part of its context.
Think of it as a permanent briefing document — one that loads before you type a single word.
Anthropic built this into Claude Code specifically for developers who work across multiple projects and need Claude to understand context without re-establishing it every session. But for marketers and agency owners, the use case is even more powerful: you’re managing multiple clients, multiple voices, multiple goals, and multiple sets of constraints simultaneously. CLAUDE.md is how you encode all of that once and get consistent output every time.
Why This Matters More for Agencies Than for Developers
A developer using CLAUDE.md might put in their tech stack, coding conventions, and repository structure. Useful, but relatively static information.
An agency marketer has far more nuanced context to pass: brand voice, competitive sensitivities, client-approved messaging, audience personas, seasonal priorities, platform preferences, things the client has explicitly rejected in the past, and the political dynamics of who approves what.
None of that lives in the code. All of it lives in your head — or in scattered Notion pages, Google Docs, and Slack threads. CLAUDE.md is how you externalize that institutional knowledge into something Claude can actually use.
Without it, you get generic outputs that need heavy editing. With it, you get first drafts that land inside the guardrails on the first try.
What to Put in a Client CLAUDE.md
Here’s the structure I use for every client. You don’t need all of these sections for every engagement — start with what’s most relevant and add to it over time.
1. Client Overview (2-3 sentences)
Who the client is, what they sell, and who buys it. Be specific. “B2B SaaS” is less useful than “B2B SaaS selling procurement automation to mid-market manufacturers with 250-2,500 employees.”
## Client Overview
Acme Logistics is a freight brokerage platform targeting regional trucking companies
with 10–100 drivers. Their buyers are operations managers and owner-operators who
distrust technology vendors and make decisions slowly. Average deal cycle: 4–6 months.
2. Brand Voice
Go beyond adjectives like “professional” or “friendly.” Give Claude concrete guidance with examples of what the voice sounds like — and equally important, what it doesn’t sound like.
## Brand Voice
- Tone: Direct, no-nonsense, practical. Like a trusted dispatcher, not a startup founder.
- Sentence length: Short. If a sentence runs over 20 words, break it.
- Vocabulary: Industry terms OK (LTL, drayage, dwell time). Avoid marketing buzzwords.
- Avoid: Exclamation points, superlatives ("best-in-class"), passive voice.
- Never use: "leverage," "seamless," "game-changer," "unlock," "empower"
3. Audience Personas
Name your primary and secondary audiences. Include what they care about most and what objections they raise. Claude will use this to calibrate argument structure, not just tone.
## Primary Audience
**Marcus, Operations Manager at a 40-truck regional carrier.**
- Pain: Brokers waste his time with low-quality loads and poor communication.
- Goal: Predictable revenue, less admin overhead.
- Objection: "We've tried software before and it just created more work."
- Responds to: Specific numbers, peer stories, low-commitment trial offers.
4. Current Priorities
What is the client actually trying to accomplish right now? This section should be updated monthly. It tells Claude what to optimize content toward.
## Current Priorities (updated March 2026)
- Primary: Drive trial signups for new Owner-Operator tier (launched Feb 2026)
- Secondary: Reduce churn in existing SMB accounts (Q1 focus)
- Do NOT push: Enterprise tier — currently paused for product reasons
5. Approved Messaging and Claims
Paste in any messaging the client has approved through legal or brand review. This prevents Claude from writing claims you’ll have to walk back.
## Approved Claims
- "Reduce empty miles by up to 23%" (backed by Q3 2025 customer study, n=47)
- "Average broker fee 12% below market rate" (internal data, do not cite publicly)
- Approved differentiators: speed of dispatch, load quality, real-time tracking
6. Competitive Sensitivities
Who are the main competitors and how should Claude handle them?
## Competitive Notes
- Main competitors: Convoy (shut down 2023 — do not reference as active), Uber Freight, Echo Global
- Do not disparage competitors by name in any client-facing content
- Positioning: We win on relationship quality and load reliability, not price
7. Platform and Format Defaults
If you’re consistently producing content for specific channels, tell Claude the defaults so you don’t have to specify them every time.
## Default Formats
- LinkedIn posts: 150-200 words, no hashtags, one line break between paragraphs
- Email subject lines: Under 45 characters, no emoji, no ALL CAPS
- Blog posts: H2 subheads every 250-300 words, include a TL;DR at the top
8. Hard Rules
Anything that must always or never happen regardless of the task.
## Hard Rules
- Always use Oxford comma
- Never promise specific ROI numbers unless pulled from Approved Claims above
- All prices in USD; do not reference Canadian pricing without confirming first
- Legal review required before any content mentioning FMCSA regulations
How to Set It Up: One Folder Per Client
The file only loads when Claude Code is running in the same directory. So the simplest structure is to give each client their own working folder:
~/clients/
acme-logistics/
CLAUDE.md ← loads automatically
briefs/
drafts/
reports/
summit-dental/
CLAUDE.md
briefs/
drafts/
When you open Claude Code inside ~/clients/acme-logistics/, it reads the Acme CLAUDE.md. Switch to ~/clients/summit-dental/ and it reads that one instead. No copy-pasting, no re-briefing, no mental overhead.
You can also nest CLAUDE.md files. A root-level one for agency-wide defaults (your billing preferences, your internal style, your standard disclaimers), and client-level ones for specific context. Claude reads both.
Building It the First Time
Don’t try to write the whole file from scratch. Instead, use this prompt inside a new Claude Code session with existing client materials (brand guide, past briefs, a few approved pieces of content):
“Read the files in this folder. Based on what you find, draft a CLAUDE.md file for this client that covers brand voice, audience, approved messaging, and content defaults. Format it in Markdown with clear section headers. Flag anything you’re uncertain about so I can fill it in.”
Claude will generate a first draft from whatever you have. You review, correct, and add the things only you know (competitive sensitivities, approved claims, political landmines). The whole process takes 15–20 minutes per client, once.
After that, you maintain it the same way you’d maintain any living document: update the Current Priorities section monthly, add new approved claims as they get signed off, remove anything that’s no longer relevant.
One More Thing: CLAUDE.md Is Not a Prompt
A common mistake is writing CLAUDE.md like a chat message or a prompt. “Please always be professional and write in a friendly tone.” Claude will technically read it, but this kind of vague instruction produces vague results.
Write it like a style guide and a briefing document combined. Be specific. Use examples. Show Claude what “good” looks like for this client, not just what you want in the abstract. The more concrete the CLAUDE.md, the closer the first draft will be to the final version.
The goal is to get to a world where you open Claude Code, type a one-line task — “write a LinkedIn post about our new Owner-Operator pricing” — and get something that’s 80% there without any additional context. That’s achievable. But it requires doing the upfront work of documenting what Claude needs to know.
CLAUDE.md is where that work lives.
What’s Next
Frequently Asked Questions
What is a CLAUDE.md file and what does it do?
A CLAUDE.md file is a plain Markdown document that Claude Code reads automatically at the start of every session in the folder where it’s saved. It acts as a permanent briefing document — loading client context, brand voice, personas, approved claims, and hard rules before you type a single prompt. The result is that Claude starts each session already knowing the client, instead of you re-establishing context from scratch every time.
Does CLAUDE.md work with Claude.ai (the website), or only Claude Code?
Only Claude Code. The auto-loading behavior — where Claude reads the file automatically when you open a session — is specific to Claude Code. In Claude.ai, you’d have to manually paste the contents of a CLAUDE.md file into the chat. If you’re using the web version, you can still use the same structure and content as a manual paste — but the automation only works inside Claude Code.
Where exactly does the CLAUDE.md file need to be saved?
It needs to be in the same directory you open Claude Code from. If your client folder is at ~/clients/acme/ and you open Claude Code inside that folder, a CLAUDE.md file at ~/clients/acme/CLAUDE.md will load automatically. Claude Code also supports nested CLAUDE.md files — a root-level one for agency-wide defaults, and client-level ones for specific context. Both load in the same session.
How long should a CLAUDE.md file be?
Long enough to be useful, short enough to stay current. Most effective CLAUDE.md files for client work run 300–600 words — enough to cover brand voice, audience, current priorities, approved claims, and hard rules without becoming a document no one maintains. The sections most worth including on day one: brand voice with examples, primary audience persona, and hard rules. Add the rest as you learn what Claude keeps getting wrong without the context.
Can multiple team members share the same CLAUDE.md file?
Yes — and this is one of the most underrated use cases for agencies. If your team stores client folders in a shared location (a synced drive, a Git repo, a shared filesystem), everyone who opens Claude Code in that client folder gets the same briefing. New team members onboard to a client in minutes instead of hours. Updates made by one person — new approved claims, updated priorities — immediately apply to every teammate’s next session.
Does CLAUDE.md affect what Claude is allowed to do, or just what it knows?
Both. CLAUDE.md can include factual context (who the client is, what they sell) and behavioral rules (never use passive voice, always include a CTA, legal review required before publishing regulatory claims). Claude treats instructions in CLAUDE.md as authoritative — it follows hard rules the same way it would follow instructions in a prompt. The key is specificity: vague guidance like “be professional” produces vague results. Specific rules like “sentences must be under 20 words” produce consistent ones.
What’s Next
Once you have CLAUDE.md set up for your clients, the next step is connecting Claude to your actual data sources — Google Search Console, Ahrefs, Notion — so it’s not just context-aware but also data-aware. That’s what MCP unlocks, and it’s what the next post covers in full, including step-by-step setup for all three tools.
Once you have MCP running, the workflows get significantly more powerful — like running a full content audit in 90 seconds by combining live GSC data with Claude’s analysis layer.
If you want to skip ahead and get the full system — CLAUDE.md templates, MCP setup guides, and the complete Claude Code workflow I use for agency work — it’s all in The AI Marketing Stack.